In this article, the problem of detecting JPEG images, which have been steganographically manipulated, is discussed. The performance of employing various shallow and deep learning algorithms in image steganography detection is analyzed. The data, images from the BOSS database, were used with information hidden using three popular steganographic algorithms: JPEG universal wavelet relative distortion (J-Uniward), nsF5, and uniform embedding revisited distortion (UERD) at two density levels. Various feature spaces were verified, with the discrete cosine transform residuals (DCTR) and the Gabor filter residuals (GFR) yielding best results. Almost perfect detection was achieved for the nsF5 algorithm at 0.4 bpnzac density (99.9% accuracy), while the detection of J-Uniward at 0.1 bpnzac density turned out to be hardly possible (max. 56.3% accuracy). The ensemble classifiers turned out to be an encouraging alternative to deep learning-based detection methods.

W tym referacie poruszamy problem wykrywania danych ukrytych steganograficznie w plikach JPEG. Zostały przedstawione najpopularniejsze algorytmy służące do ukrywania informacji w obrazach, a następnie zostały omówione algorytmy służące do ekstrakcji cech z obrazów cyfrowych. Pokazane zostały metody opisane w literaturze bazujące na ekstrakcji cech DCTR, GFR oraz PHARM oraz sieciach neuronowych w architekturze Convolution-Batch- Normalization-Dense. Zostało zaproponowane nowe rozwiązanie z wykorzystaniem prostej sieci neuronowej, przeprowadzone zostały badania dokładności oraz innych metryk dla najlepszej konfiguracji sieci.

The purpose of the paper is to compare the computational complexity of the steganalytic method created by the authors in previously developed and presented by other academics to detect three steganographic algorithms, including jUniward, nsF5 and UERD, against other solutions found during the literature survey. The article describes shortly each analyzed steganographic algorithm along with all compared solutions, gives their characteristic points, and includes list of the model specifications. Moreover, the metric comparing the computational complexity, so the number of parameters in a given neural network, is presented. Since none of the compared articles provided this parameter, there is provided algorithm by which these values could be calculated for each analyzed model. The comparison of the results and final conclusions are presented to explain the distinctions between the obtained outcomes and to identify the factors which could have affected such performance.